How does TargetDiscovery handle dichotomous outcomes?
Targetdiscovery estimates the variable importance of a target variable on the mean outcome on an additive scale, i.e. it estimates the difference in the mean outcome associated with comparing different values of the target variable, adjusted for confounders. In the context of a dichotomous outcome, the mean outcome represents the probability of the outcome being equal to 1. The use of a linear variable importance model means that TargetDiscovery does not constrain the estimated mean outcome, given the target variable and confounders, to lie between 0 and 1, as would be the case if a logistic model were used, for example. Note, however, that a linear variable importance model will still provide a valid test of the null hypothesis of interest (zero variable importance). Since such tests are the foundation of a discovery analysis, the linear model is used by TargetDiscovery for both continuous and dichotomous outcomes.
